diff --git a/rag/llm/chat_model.py b/rag/llm/chat_model.py
index 95d86c4b93..6571f11701 100644
--- a/rag/llm/chat_model.py
+++ b/rag/llm/chat_model.py
@@ -61,6 +61,47 @@ ERROR_PREFIX = "**ERROR**"
 LENGTH_NOTIFICATION_CN = "······\n由于大模型的上下文窗口大小限制，回答已经被大模型截断。"
 LENGTH_NOTIFICATION_EN = "...\nThe answer is truncated by your chosen LLM due to its limitation on context length."
 
+# Generation parameters that are safe to forward to the underlying completion
+# call. `gen_conf` originates from a chat assistant's `llm_setting`, which can
+# also carry RAGFlow-internal metadata (e.g. `model_type`). Anything outside
+# this set is dropped so providers don't reject the request with errors like
+# "Extra inputs are not permitted" / "Unknown parameter: 'model_type'" (#15427).
+ALLOWED_GEN_CONF_KEYS = frozenset(
+    {
+        "temperature",
+        "max_completion_tokens",
+        "top_p",
+        "stream",
+        "stream_options",
+        "stop",
+        "n",
+        "presence_penalty",
+        "frequency_penalty",
+        "functions",
+        "function_call",
+        "logit_bias",
+        "user",
+        "response_format",
+        "seed",
+        "tools",
+        "tool_choice",
+        "logprobs",
+        "top_logprobs",
+        "extra_headers",
+    }
+)
+
+# LiteLLM additionally understands reasoning-control parameters that the
+# model-family policies may inject into `gen_conf` (e.g. `thinking` for
+# Anthropic / Kimi reasoning models, `reasoning_effort` for OpenAI o-series).
+LITELLM_ALLOWED_GEN_CONF_KEYS = ALLOWED_GEN_CONF_KEYS | frozenset(
+    {
+        "thinking",
+        "reasoning_effort",
+        "extra_body",
+    }
+)
+
 
 def _apply_model_family_policies(
     model_name: str,
@@ -159,30 +200,7 @@ class Base(ABC):
         if "max_tokens" in gen_conf:
             del gen_conf["max_tokens"]
 
-        allowed_conf = {
-            "temperature",
-            "max_completion_tokens",
-            "top_p",
-            "stream",
-            "stream_options",
-            "stop",
-            "n",
-            "presence_penalty",
-            "frequency_penalty",
-            "functions",
-            "function_call",
-            "logit_bias",
-            "user",
-            "response_format",
-            "seed",
-            "tools",
-            "tool_choice",
-            "logprobs",
-            "top_logprobs",
-            "extra_headers",
-        }
-
-        gen_conf = {k: v for k, v in gen_conf.items() if k in allowed_conf}
+        gen_conf = {k: v for k, v in gen_conf.items() if k in ALLOWED_GEN_CONF_KEYS}
         return gen_conf
 
     async def _async_chat_streamly(self, history, gen_conf, **kwargs):
@@ -1395,6 +1413,7 @@ class LiteLLMBase(ABC):
         )
 
         gen_conf.pop("max_tokens", None)
+        gen_conf = {k: v for k, v in gen_conf.items() if k in LITELLM_ALLOWED_GEN_CONF_KEYS}
         return gen_conf
 
     def _need_reasoning_content_back(self) -> bool:
diff --git a/test/unit_test/rag/graphrag/conftest.py b/test/unit_test/rag/graphrag/conftest.py
index 8aa4f43e81..a980592b5d 100644
--- a/test/unit_test/rag/graphrag/conftest.py
+++ b/test/unit_test/rag/graphrag/conftest.py
@@ -47,4 +47,12 @@ sys.modules["common.connection_utils"].timeout = lambda *a, **kw: (lambda fn: fn
 sys.modules["api.db.services.task_service"].has_canceled = lambda *_a, **_kw: False
 sys.modules["rag.graphrag.general.leiden"].run = lambda *_a, **_kw: {}
 sys.modules["rag.graphrag.general.leiden"].add_community_info2graph = lambda *_a, **_kw: None
-sys.modules["rag.llm.chat_model"].Base = object
+# Only stub ``Base`` when we actually mocked chat_model. This conftest mutates
+# the global sys.modules at import time, and rag/graphrag/ is collected before
+# rag/llm/. If an earlier test package already imported the real chat_model,
+# unconditionally assigning ``Base = object`` clobbered the genuine class and
+# leaked into the rag/llm unit tests that import it (AttributeError: no
+# attribute '_clean_conf'). graphrag only uses ``Base`` as a type alias, so the
+# real class works just as well when it is already loaded.
+if isinstance(sys.modules["rag.llm.chat_model"], MagicMock):
+    sys.modules["rag.llm.chat_model"].Base = object
diff --git a/test/unit_test/rag/llm/test_clean_conf_whitelist.py b/test/unit_test/rag/llm/test_clean_conf_whitelist.py
new file mode 100644
index 0000000000..019a27be1a
--- /dev/null
+++ b/test/unit_test/rag/llm/test_clean_conf_whitelist.py
@@ -0,0 +1,123 @@
+#
+#  Copyright 2025 The InfiniFlow Authors. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#
+"""
+Regression guard for issue #15427: LiteLLM-routed chats failed with
+``model_type: Extra inputs are not permitted`` (Anthropic) /
+``Unknown parameter: 'model_type'`` (OpenAI).
+
+A chat assistant's ``llm_setting`` is forwarded to the provider as
+``gen_conf``. ``llm_setting`` can legitimately carry RAGFlow-internal
+metadata such as ``model_type`` (the new chat REST APIs read it back out).
+``Base._clean_conf`` already whitelisted the keys it forwards, so OpenAI-
+compatible providers were unaffected, but ``LiteLLMBase._clean_conf`` only
+dropped ``max_tokens`` and passed everything else straight through to
+``litellm.acompletion`` — which forwarded ``model_type`` to the upstream
+provider and got rejected.
+
+These tests pin the whitelisting behaviour for both backends so the leak
+cannot reappear.
+"""
+
+import pytest
+
+from rag.llm.chat_model import (
+    ALLOWED_GEN_CONF_KEYS,
+    LITELLM_ALLOWED_GEN_CONF_KEYS,
+    Base,
+    LiteLLMBase,
+)
+
+
+class _ConcreteBase(Base):
+    """Concrete subclass so we can build an instance without touching the
+    real OpenAI client constructor."""
+
+
+class _ConcreteLiteLLM(LiteLLMBase):
+    """Concrete subclass for the same reason on the LiteLLM path."""
+
+
+def _make_base(model_name="gpt-4o"):
+    inst = _ConcreteBase.__new__(_ConcreteBase)
+    inst.model_name = model_name
+    return inst
+
+
+def _make_litellm(model_name="gpt-4o", provider=""):
+    inst = _ConcreteLiteLLM.__new__(_ConcreteLiteLLM)
+    inst.model_name = model_name
+    inst.provider = provider
+    return inst
+
+
+# --------------------------------------------------------------------------- #
+# The actual bug: model_type must never reach the provider.
+# --------------------------------------------------------------------------- #
+def test_litellm_drops_model_type():
+    cleaned = _make_litellm()._clean_conf({"temperature": 0.5, "model_type": "chat"})
+    assert "model_type" not in cleaned
+    assert cleaned["temperature"] == 0.5
+
+
+def test_base_drops_model_type():
+    cleaned = _make_base()._clean_conf({"temperature": 0.5, "model_type": "chat"})
+    assert "model_type" not in cleaned
+    assert cleaned["temperature"] == 0.5
+
+
+@pytest.mark.parametrize("stray_key", ["model_type", "llm_id", "parameter", "icon", "foo"])
+def test_litellm_drops_arbitrary_internal_keys(stray_key):
+    cleaned = _make_litellm()._clean_conf({stray_key: "x", "top_p": 0.9})
+    assert stray_key not in cleaned
+    assert cleaned["top_p"] == 0.9
+
+
+# --------------------------------------------------------------------------- #
+# The fix must not over-filter: genuine generation params still pass through.
+# --------------------------------------------------------------------------- #
+def test_litellm_preserves_known_generation_params():
+    gen_conf = {
+        "temperature": 0.7,
+        "top_p": 0.95,
+        "presence_penalty": 0.1,
+        "frequency_penalty": 0.2,
+    }
+    cleaned = _make_litellm()._clean_conf(dict(gen_conf))
+    assert cleaned == gen_conf
+
+
+def test_litellm_preserves_thinking_param():
+    """``thinking`` is injected by the model-family policy for reasoning
+    models and must survive the whitelist (it is a valid LiteLLM param)."""
+    cleaned = _make_litellm()._clean_conf({"thinking": {"type": "enabled"}, "temperature": 1.0})
+    assert cleaned["thinking"] == {"type": "enabled"}
+
+
+def test_max_tokens_is_dropped_on_both_backends():
+    assert "max_tokens" not in _make_litellm()._clean_conf({"max_tokens": 100, "temperature": 0.3})
+    assert "max_tokens" not in _make_base()._clean_conf({"max_tokens": 100, "temperature": 0.3})
+
+
+# --------------------------------------------------------------------------- #
+# Whitelist invariants.
+# --------------------------------------------------------------------------- #
+def test_litellm_whitelist_is_superset_of_base():
+    assert ALLOWED_GEN_CONF_KEYS <= LITELLM_ALLOWED_GEN_CONF_KEYS
+
+
+def test_model_type_not_whitelisted_anywhere():
+    assert "model_type" not in ALLOWED_GEN_CONF_KEYS
+    assert "model_type" not in LITELLM_ALLOWED_GEN_CONF_KEYS